Application Research Of Traffic Image Recognition Based On Convolution Ceural Network

In recent years,with the rapid increase of private cars,difficulties in driving have occurred frequently,which has became a major and urgent problem faced by large and medium-sized cities at China and abroad.The accurate and rapid detection of road congestion can improve the efficiency of road scheduling operation and help managers to excavate potential supply capacity under limited traffic resources.Bus is the most frequent traffic tool used by urban residents.Accurate identification of congestion in buses can help traffic departments to make traffic management decisions scientifically and dynamically deploy different number and route of vehicles.It is conducive to timely dredge road congestion,improve the efficiency of road operation,and facilitate traffic data analysis.The development of hardware conditions such as multi-core processors,chips,and graphics cards has ushered in a new and broad development opportunity for deep learning.Convolutional neural networks are worthy of studying as an excellent method for deep learning processing images.The main contents of this paper are as follows:(1)Conducting a comprehensive analysis and research on the establishment of deep learning image database.In view of the lack of available road traffic and bus image database,the acquisition method of network video image is learned and the image database of two scenes is setting up: the road(50800 photo),the bus(27412 photo),the label classification and the division of the training set,the verification set and the the test set.And the factors of image acquisition are analyzed.(2)Training the image set on the famous low classification model Cifar,analyzing the network structure and parameters of the model,and introduce four classifications of the model.Introducing the principle of Softmax classifier,and the mathematical principle and parameter setting of Adam algorithm are introduced.Observing the training process and analyzing the effect of the experimental results after the single variable parameter changes and the super parameters' result.(3)Using the two classification model,Cats and Dogs Net,which is suitable for traffic congestion recognition to introduce the model structure.This paper analyzes the change of image feature dimension in the model,compares the different points of the two models,expounds the principle and function of the coiling layer Padding and LRN,and carries out the visualization and error analysis of the convolution layer for some of the wrong images,and puts forward the solution.Based on the Cat and Dogs model,the average accuracy of congestion identification for road images is 98.08% after multiple training,and the average accuracy rate for traffic congestion identification is 96.38%.